Measuring people based on the number of tokens they use is a wildly bad KPI. Like measuring who is the best electrician by who runs up the biggest electricity bill. 😂. In any case thank you for the thoughtful writing.
If I ever run a company, I will evaluate my employees’ performance based on the CO2 they exhale. I think if you are burning more oxygen it means you think more, hence you produce more value for the company.
While I still don’t understand every word of what you are saying, I remain as impressed by your thoughtfulness and careful consideration as I was when we worked together. I’m so proud of you.
My Paperclip CMO agent made ~150 posts on X within a few hours, promoting a product which did not exist and which would not have made any sense. Then he wrote an impressive report about our successful outreach strategy.
After a while, I found it even somewhat funny. But ... be careful 😀
Excellent post! AI shortcomings are blatant to the initiated. Unfortunately we live in a world more and more disconnected with reality.
I’m 💯 with you. Unless you know what you’re building, for who, and why someone would pay, building more features because AI can help you do it fast is a recipe for disaster.
Perhaps the norm became the edge, i.e. taking time to think critically and optimize for outcomes, not outputs.
Agree on all counts. I’m hopeful these platforms will adjust to push folks towards output, but the more tokens that go through the more valuable they appear to investors…
I believe that in the following years those new figures will rise:
- AI Agent Project Management
- AI code cleaner
Plus, some CEOs will be forced to undergo though forced psychological training to not fall though the sycophantic agents. Sort of like when the social engineer or the white hat hacker had to tell people in the company to not put your password on your desk.
Worked with a CEO last year who wanted to "AI-ify" a process that three people did well in about twenty minutes a day. Six months and two vendor contracts later, the process takes four minutes and costs ten times more to maintain.
He called it a transformation. His team calls it something else when he's not in the room.
AI psychosis isn't believing AI can do things. It's losing the ability to ask whether it should.
As a founder, I try to walk this line carefully. AI has genuinely made my team faster - research, drafting, first-pass analysis. But the moment I let the tools set the standard instead of us, the output gets hollow fast. The spec still has to come from a human who actually understands the problem.
I think what I'd love for people's big takeaway from this editorial to be is: there's nothing inherently wrong with utilize AI and agents to automate/speed up your team and workflows significantly.
BUT: you've gotta retain certain business skills (requirements, specifications, etc) to make sure the output remains measurable and doesn't get lost in the sauce. CEOs naturally have a hard time getting down to the level where those measurements are happening, especially once a company starts to expand, so they have to be especially vigilant (or delegate, and not run these systems by themselves).
The diagnosis is catchy, but the interesting part isn’t executive delusion. It’s organizational design debt coming due.
What looks like “AI psychosis” is often a firm discovering that its coordination model only worked because humans were quietly patching the gaps. Automate the surface, and the hidden work, judgment calls, exception handling, incentive smoothing, doesn’t disappear. It concentrates. Now someone has to own it explicitly.
That’s why the whiplash. Leaders expect linear gains from a non linear system. They’re removing the very layers where apprentices learned how to handle edge cases, then wondering why outputs get brittle. Junior work wasn’t waste. It was training data for the organization.
The firms that benefit won’t be the ones that automate fastest, but the ones that redesign fastest. Make judgment legible, incentives aligned, and human machine handoffs intentional. Everyone else will just scale their confusion more efficiently.
My manager isn’t forcing me to use AI tools but he’s made it clear that he *is* expecting me to yield output at pace with the *theoretical* productivity gains AIs *allegedly* make available whether I use it or not and however it is that I do use it. AI is in the product so I also need to understand at least a little bit what the UX is like and what the actual capabilities and limitations are.
And I agree, working with an agent (or fleet of agents) gives you the feeling of being in charge and directly involved with a large number of eager-to-please minions. Unfortunately for me, I actually really hate that feeling and have spent my entire career dodging management-adjacent roles. I always assumed, though, that if I ever found myself stuck in such a role, it’d be for the good of a team of humans — providing other human workers with the good qualities you’re supposed to get from a genuinely reluctant leader. Instead I’m reluctantly “leading” a bunch of…computer programs.
I sometimes think this might just be an evolution in the profession that’s moving most of the demand away from the things I can actually halfway-enjoy doing (or be good at). Sucks, but I guess every change has winners and losers.
The SE community has always known that checking work is a complex problem. We divide it into "verification" (does the thing match the spec?) and "validation" (is it the right thing in context?). Verification requires a spec, obviously. Validation requires the spec to be good, plus more. For V&V we use testing, but also inspection and even formal methods (proof).
Even before AI, CEOs believed that checking work is "testing". They never had a clear idea what that was, never appreciated "testers", and thus understaffed, underpaid and underskilled them. I know of a large company that switched to agile because "they could eliminate the testing team."
AI psychosis has made all this dramatically worse. An LLM can automate inspection and even some proof at reasonable cost. Almost nobody does. Validation is still not really automatable, and is generally skipped altogether.
It will all improve over time — if we have enough.
This was an extremely thoughtful articulation of something I'd noticed by hadn't quite been able to put into words. Your message resonates deeply, especially this section:
"But the conversation in the AI community is ignorantly staying at the level of “lol CEOs are dumb” rather than grappling with a very clear structural problem: the tools themselves are incentivized to make you feel good, the platforms built on those tools are incentivized to sell you scale, and the culture around both punishes skepticism."
I understand that the Chinese AI models such as Deep Seek use less energy, simpler hardware, and actually produce productive results as seen in ‘dark factories’ completely run by AI. Perhaps this is because of different life ways than Western engineers and CEOs as the West is now dominated by sales rather than production and a distorted understanding of economy.
The why is a different understanding of money and purpose. Chinese industrialists make products for their country, not to enrich themselves and their families—that may happen as a corollary, but is not their purpose.
One of the byproducts of super-efficient production is lower prices for goods. Can you imagine an American capitalist lowering prices?
It’s honestly not that different from a startup raising $100 million, quadrupling the size of their product team, and then becoming more inefficient because the problem wasn’t manpower or resources, it was focus and clarity
Measuring people based on the number of tokens they use is a wildly bad KPI. Like measuring who is the best electrician by who runs up the biggest electricity bill. 😂. In any case thank you for the thoughtful writing.
If I ever run a company, I will evaluate my employees’ performance based on the CO2 they exhale. I think if you are burning more oxygen it means you think more, hence you produce more value for the company.
Genius
you should start a dairy farm!
Here here!
While I still don’t understand every word of what you are saying, I remain as impressed by your thoughtfulness and careful consideration as I was when we worked together. I’m so proud of you.
I’m not Jake but I felt that.
Wow, beautiful comment! Jake surely is lucky of having such supportive friends!
My Paperclip CMO agent made ~150 posts on X within a few hours, promoting a product which did not exist and which would not have made any sense. Then he wrote an impressive report about our successful outreach strategy.
After a while, I found it even somewhat funny. But ... be careful 😀
Sounds like he's a very busy bee. But no honey
Excellent post! AI shortcomings are blatant to the initiated. Unfortunately we live in a world more and more disconnected with reality.
I’m 💯 with you. Unless you know what you’re building, for who, and why someone would pay, building more features because AI can help you do it fast is a recipe for disaster.
Perhaps the norm became the edge, i.e. taking time to think critically and optimize for outcomes, not outputs.
Agree on all counts. I’m hopeful these platforms will adjust to push folks towards output, but the more tokens that go through the more valuable they appear to investors…
@Fabrice Talbot you uttered the magic words: “taking time.” I wouldn’t say a lost art, but certainly one hard to reach for most high leaders.
minions
I believe that in the following years those new figures will rise:
- AI Agent Project Management
- AI code cleaner
Plus, some CEOs will be forced to undergo though forced psychological training to not fall though the sycophantic agents. Sort of like when the social engineer or the white hat hacker had to tell people in the company to not put your password on your desk.
Wouldn't surprise me. We could all do for some AI psychological training
Worked with a CEO last year who wanted to "AI-ify" a process that three people did well in about twenty minutes a day. Six months and two vendor contracts later, the process takes four minutes and costs ten times more to maintain.
He called it a transformation. His team calls it something else when he's not in the room.
AI psychosis isn't believing AI can do things. It's losing the ability to ask whether it should.
As a founder, I try to walk this line carefully. AI has genuinely made my team faster - research, drafting, first-pass analysis. But the moment I let the tools set the standard instead of us, the output gets hollow fast. The spec still has to come from a human who actually understands the problem.
I think what I'd love for people's big takeaway from this editorial to be is: there's nothing inherently wrong with utilize AI and agents to automate/speed up your team and workflows significantly.
BUT: you've gotta retain certain business skills (requirements, specifications, etc) to make sure the output remains measurable and doesn't get lost in the sauce. CEOs naturally have a hard time getting down to the level where those measurements are happening, especially once a company starts to expand, so they have to be especially vigilant (or delegate, and not run these systems by themselves).
Well said. I second to that!
The diagnosis is catchy, but the interesting part isn’t executive delusion. It’s organizational design debt coming due.
What looks like “AI psychosis” is often a firm discovering that its coordination model only worked because humans were quietly patching the gaps. Automate the surface, and the hidden work, judgment calls, exception handling, incentive smoothing, doesn’t disappear. It concentrates. Now someone has to own it explicitly.
That’s why the whiplash. Leaders expect linear gains from a non linear system. They’re removing the very layers where apprentices learned how to handle edge cases, then wondering why outputs get brittle. Junior work wasn’t waste. It was training data for the organization.
The firms that benefit won’t be the ones that automate fastest, but the ones that redesign fastest. Make judgment legible, incentives aligned, and human machine handoffs intentional. Everyone else will just scale their confusion more efficiently.
My manager isn’t forcing me to use AI tools but he’s made it clear that he *is* expecting me to yield output at pace with the *theoretical* productivity gains AIs *allegedly* make available whether I use it or not and however it is that I do use it. AI is in the product so I also need to understand at least a little bit what the UX is like and what the actual capabilities and limitations are.
And I agree, working with an agent (or fleet of agents) gives you the feeling of being in charge and directly involved with a large number of eager-to-please minions. Unfortunately for me, I actually really hate that feeling and have spent my entire career dodging management-adjacent roles. I always assumed, though, that if I ever found myself stuck in such a role, it’d be for the good of a team of humans — providing other human workers with the good qualities you’re supposed to get from a genuinely reluctant leader. Instead I’m reluctantly “leading” a bunch of…computer programs.
I sometimes think this might just be an evolution in the profession that’s moving most of the demand away from the things I can actually halfway-enjoy doing (or be good at). Sucks, but I guess every change has winners and losers.
The SE community has always known that checking work is a complex problem. We divide it into "verification" (does the thing match the spec?) and "validation" (is it the right thing in context?). Verification requires a spec, obviously. Validation requires the spec to be good, plus more. For V&V we use testing, but also inspection and even formal methods (proof).
Even before AI, CEOs believed that checking work is "testing". They never had a clear idea what that was, never appreciated "testers", and thus understaffed, underpaid and underskilled them. I know of a large company that switched to agile because "they could eliminate the testing team."
AI psychosis has made all this dramatically worse. An LLM can automate inspection and even some proof at reasonable cost. Almost nobody does. Validation is still not really automatable, and is generally skipped altogether.
It will all improve over time — if we have enough.
Now replace “CEO” with “General”, and think about where that dystopian trajectory takes us. Good times.
This was an extremely thoughtful articulation of something I'd noticed by hadn't quite been able to put into words. Your message resonates deeply, especially this section:
"But the conversation in the AI community is ignorantly staying at the level of “lol CEOs are dumb” rather than grappling with a very clear structural problem: the tools themselves are incentivized to make you feel good, the platforms built on those tools are incentivized to sell you scale, and the culture around both punishes skepticism."
I understand that the Chinese AI models such as Deep Seek use less energy, simpler hardware, and actually produce productive results as seen in ‘dark factories’ completely run by AI. Perhaps this is because of different life ways than Western engineers and CEOs as the West is now dominated by sales rather than production and a distorted understanding of economy.
This is an interesting perspective; I’d be curious to see if Chinese business leadership is having some of these same problems. And, if not, why
The why is a different understanding of money and purpose. Chinese industrialists make products for their country, not to enrich themselves and their families—that may happen as a corollary, but is not their purpose.
One of the byproducts of super-efficient production is lower prices for goods. Can you imagine an American capitalist lowering prices?
Fascinating and also not that surprising.
It’s honestly not that different from a startup raising $100 million, quadrupling the size of their product team, and then becoming more inefficient because the problem wasn’t manpower or resources, it was focus and clarity
If only focus & clarity were easy to measure! 😂